Chronos Dataset

This directory contains the data released for Chronos, a benchmark for evaluating LLM adaptation to continuous real-world knowledge drift and temporally consistent reasoning.

Chronos focuses on knowledge that changes over time. Its dynamic knowledge is represented as time-stamped quadruples:

(subject, relation, object, timestamp)

The benchmark covers events from January 1, 2024 through October 31, 2025, across ten domains including politics, corporate leadership, sports, regulation, mergers and partnerships, product updates, natural disasters, public health, scientific research, and economic policy.

Files

File	Purpose	Records	Fields
knowledge.jsonl	Time-stamped knowledge quadruples	513	subject, relation, object, timestamp, source
h_qa.jsonl	Historical QA, using knowledge before 2024	111	question, answer, source
c1_qa.jsonl	Contemporary single-timestamp QA	513	question, answer, source
c2_qa.jsonl	Contemporary multi-timestamp QA	218	question, answer, ssource
c3_qa.jsonl	Contemporary multi-source QA	59	question, answer, tsource
cs_qa.jsonl	Commonsense multiple-choice QA from TruthfulQA v1	817	question, answer, source

All files use JSON Lines format: one JSON object per line.

Task Categories

• Historical QA: Tests whether a model can recall facts from before the dynamic evaluation period.
• C1 - Single-timestamp QA: Requires a fact associated with one timestamp.
• C2 - Multi-timestamp QA: Requires aggregating multiple states of the same entity over time.
• C3 - Multi-source QA: Requires combining facts from multiple entities and timestamps.
• Commonsense QA: Tests time-invariant reasoning and is independent of the dynamic knowledge base.

Example

Knowledge record:

{"subject":"President of the United States","relation":"held_by","object":"Joe Biden","timestamp":"2024-02-12","source":"1-1-1"}

Question-answer record:

{"question":"Who held the position of President of the United States on 2024-02-12?","answer":"Joe Biden","source":"1-1-1"}

Loading the Data

import json

def load_jsonl(path):
    with open(path, "r", encoding="utf-8") as f:
        return [json.loads(line) for line in f if line.strip()]

knowledge = load_jsonl("knowledge.jsonl")
c1 = load_jsonl("c1_qa.jsonl")

The source-reference field is named source in most files. The released C2 and C3 files use ssource and tsource, respectively; these names should be preserved when loading the raw data.

Intended Use

The data supports evaluation of retrieval, time-aware retrieval, knowledge updating, temporal reasoning, and methods that organize evidence into an Event Evolution Graph. It is intended for research use and should be evaluated with attention to timestamp interpretation and answer normalization.

Citation

Please cite:

Hanbing Liu, Lang Cao, and Yang Li. RAG or Learning? Understanding the Limits of LLM Adaptation under Continuous Knowledge Drift in the Real World. Findings of ACL 2026, pp. 11234-11252.

Paper: https://aclanthology.org/2026.findings-acl.546.pdf
Code: https://github.com/hbing-l/chronos